2 % (c) The University of Glasgow 2000
4 \section[ByteCodeGen]{Generate bytecode from Core}
7 module ByteCodeGen ( UnlinkedBCO, UnlinkedBCOExpr, ItblEnv, ClosureEnv, HValue,
9 byteCodeGen, coreExprToBCOs
12 #include "HsVersions.h"
15 import Name ( Name, getName )
16 import Id ( Id, idType, isDataConId_maybe, isPrimOpId_maybe, isFCallId,
17 idPrimRep, mkSysLocal, idName, isFCallId_maybe, isPrimOpId )
18 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
19 import OrdList ( OrdList, consOL, snocOL, appOL, unitOL,
20 nilOL, toOL, concatOL, fromOL )
21 import FiniteMap ( FiniteMap, addListToFM, listToFM, elemFM,
22 addToFM, lookupFM, fmToList )
24 import PprCore ( pprCoreExpr )
25 import Literal ( Literal(..), literalPrimRep )
26 import PrimRep ( PrimRep(..) )
27 import PrimOp ( PrimOp(..) )
28 import CoreFVs ( freeVars )
29 import Type ( typePrimRep, splitTyConApp_maybe, isTyVarTy )
30 import DataCon ( dataConTag, fIRST_TAG, dataConTyCon,
31 dataConWrapId, isUnboxedTupleCon )
32 import TyCon ( TyCon(..), tyConFamilySize, isDataTyCon, tyConDataCons,
33 isFunTyCon, isUnboxedTupleTyCon )
34 import Class ( Class, classTyCon )
35 import Type ( Type, repType, splitRepFunTys )
36 import Util ( zipEqual, zipWith4Equal, naturalMergeSortLe, nOfThem,
37 isSingleton, lengthIs )
38 import DataCon ( dataConRepArity )
39 import Var ( isTyVar )
40 import VarSet ( VarSet, varSetElems )
41 import PrimRep ( isFollowableRep )
42 import CmdLineOpts ( DynFlags, DynFlag(..) )
43 import ErrUtils ( showPass, dumpIfSet_dyn )
44 import Unique ( mkPseudoUnique3 )
45 import FastString ( FastString(..) )
46 import Panic ( GhcException(..) )
47 import PprType ( pprType )
48 import SMRep ( arrWordsHdrSize, arrPtrsHdrSize )
49 import Constants ( wORD_SIZE )
50 import ByteCodeInstr ( BCInstr(..), ProtoBCO(..), nameOfProtoBCO, bciStackUse )
51 import ByteCodeItbls ( ItblEnv, mkITbls )
52 import ByteCodeLink ( UnlinkedBCO, UnlinkedBCOExpr, assembleBCO,
53 ClosureEnv, HValue, filterNameMap, linkFail,
54 iNTERP_STACK_CHECK_THRESH )
55 import ByteCodeFFI ( taggedSizeW, untaggedSizeW, mkMarshalCode, moan64 )
56 import Linker ( lookupSymbol )
58 import List ( intersperse, sortBy, zip4 )
59 import Foreign ( Ptr(..), castPtr, mallocBytes, pokeByteOff, Word8 )
60 import CTypes ( CInt )
61 import Exception ( throwDyn )
63 import GlaExts ( Int(..), ByteArray# )
66 import Maybe ( isJust )
70 %************************************************************************
72 \subsection{Functions visible from outside this module.}
74 %************************************************************************
78 byteCodeGen :: DynFlags
81 -> IO ([UnlinkedBCO], ItblEnv)
82 byteCodeGen dflags binds local_tycons local_classes
83 = do showPass dflags "ByteCodeGen"
84 let tycs = local_tycons ++ map classTyCon local_classes
85 itblenv <- mkITbls tycs
87 let flatBinds = concatMap getBind binds
88 getBind (NonRec bndr rhs) = [(bndr, freeVars rhs)]
89 getBind (Rec binds) = [(bndr, freeVars rhs) | (bndr,rhs) <- binds]
91 (BcM_State proto_bcos final_ctr mallocd, ())
92 <- runBc (BcM_State [] 0 [])
93 (mapBc (schemeR True []) flatBinds `thenBc_` returnBc ())
95 -- better be no free vars in these top-level bindings
97 when (not (null mallocd))
98 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
100 dumpIfSet_dyn dflags Opt_D_dump_BCOs
101 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
103 bcos <- mapM assembleBCO proto_bcos
105 return (bcos, itblenv)
108 -- Returns: (the root BCO for this expression,
109 -- a list of auxilary BCOs resulting from compiling closures)
110 coreExprToBCOs :: DynFlags
112 -> IO UnlinkedBCOExpr
113 coreExprToBCOs dflags expr
114 = do showPass dflags "ByteCodeGen"
116 -- create a totally bogus name for the top-level BCO; this
117 -- should be harmless, since it's never used for anything
118 let invented_id = mkSysLocal FSLIT("Expr-Top-Level")
120 (panic "invented_id's type")
121 let invented_name = idName invented_id
123 annexpr = freeVars expr
124 fvs = filter (not.isTyVar) (varSetElems (fst annexpr))
126 (BcM_State all_proto_bcos final_ctr mallocd, ())
127 <- runBc (BcM_State [] 0 [])
128 (schemeR True fvs (invented_id, annexpr))
130 when (not (null mallocd))
131 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
133 dumpIfSet_dyn dflags Opt_D_dump_BCOs
134 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr all_proto_bcos)))
137 = case filter ((== invented_name).nameOfProtoBCO) all_proto_bcos of
138 [root_bco] -> root_bco
140 = filter ((/= invented_name).nameOfProtoBCO) all_proto_bcos
142 auxiliary_bcos <- mapM assembleBCO auxiliary_proto_bcos
143 root_bco <- assembleBCO root_proto_bco
145 return (root_bco, auxiliary_bcos)
148 %************************************************************************
150 \subsection{Compilation schema for the bytecode generator.}
152 %************************************************************************
156 type BCInstrList = OrdList BCInstr
158 type Sequel = Int -- back off to this depth before ENTER
160 -- Maps Ids to the offset from the stack _base_ so we don't have
161 -- to mess with it after each push/pop.
162 type BCEnv = FiniteMap Id Int -- To find vars on the stack
164 ppBCEnv :: BCEnv -> SDoc
167 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
170 pp_one (var, offset) = int offset <> colon <+> ppr var
171 cmp_snd x y = compare (snd x) (snd y)
173 -- Create a BCO and do a spot of peephole optimisation on the insns
175 mkProtoBCO nm instrs_ordlist origin mallocd_blocks
176 = ProtoBCO nm maybe_with_stack_check origin mallocd_blocks
178 -- Overestimate the stack usage (in words) of this BCO,
179 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
180 -- stack check. (The interpreter always does a stack check
181 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
182 -- BCO anyway, so we only need to add an explicit on in the
183 -- (hopefully rare) cases when the (overestimated) stack use
184 -- exceeds iNTERP_STACK_CHECK_THRESH.
185 maybe_with_stack_check
186 | stack_overest >= 65535
187 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
189 | stack_overest >= iNTERP_STACK_CHECK_THRESH
190 = (STKCHECK stack_overest) : peep_d
192 = peep_d -- the supposedly common case
194 stack_overest = sum (map bciStackUse peep_d)
195 + 10 {- just to be really really sure -}
198 -- Merge local pushes
199 peep_d = peep (fromOL instrs_ordlist)
201 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
202 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
203 peep (PUSH_L off1 : PUSH_L off2 : rest)
204 = PUSH_LL off1 (off2-1) : peep rest
211 -- Compile code for the right hand side of a let binding.
212 -- Park the resulting BCO in the monad. Also requires the
213 -- variable to which this value was bound, so as to give the
214 -- resulting BCO a name. Bool indicates top-levelness.
216 schemeR :: Bool -> [Id] -> (Id, AnnExpr Id VarSet) -> BcM ()
217 schemeR is_top fvs (nm, rhs)
221 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
222 $$ pprCoreExpr (deAnnotate rhs)
228 = schemeR_wrk is_top fvs rhs nm (collect [] rhs)
231 collect xs (_, AnnNote note e)
233 collect xs (_, AnnLam x e)
234 = collect (if isTyVar x then xs else (x:xs)) e
235 collect xs not_lambda
236 = (reverse xs, not_lambda)
238 schemeR_wrk is_top fvs original_body nm (args, body)
239 | Just dcon <- maybe_toplevel_null_con_rhs
240 = --trace ("nullary constructor! " ++ showSDocDebug (ppr nm)) (
241 emitBc (mkProtoBCO (getName nm) (toOL [PACK dcon 0, ENTER])
242 (Right original_body))
246 = let all_args = reverse args ++ fvs
247 szsw_args = map taggedIdSizeW all_args
248 szw_args = sum szsw_args
249 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
250 argcheck = unitOL (ARGCHECK szw_args)
252 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
253 emitBc (mkProtoBCO (getName nm) (appOL argcheck body_code)
254 (Right original_body))
257 maybe_toplevel_null_con_rhs
258 | is_top && null args
259 = case nukeTyArgs (snd body) of
261 -> case isDataConId_maybe v_wrk of
263 Just dc_wrk | nm == dataConWrapId dc_wrk
271 nukeTyArgs (AnnApp f (_, AnnType _)) = nukeTyArgs (snd f)
272 nukeTyArgs other = other
275 -- Let szsw be the sizes in words of some items pushed onto the stack,
276 -- which has initial depth d'. Return the values which the stack environment
277 -- should map these items to.
278 mkStackOffsets :: Int -> [Int] -> [Int]
279 mkStackOffsets original_depth szsw
280 = map (subtract 1) (tail (scanl (+) original_depth szsw))
282 -- Compile code to apply the given expression to the remaining args
283 -- on the stack, returning a HNF.
284 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr Id VarSet -> BcM BCInstrList
286 -- Delegate tail-calls to schemeT.
287 schemeE d s p e@(fvs, AnnApp f a)
288 = schemeT d s p (fvs, AnnApp f a)
290 schemeE d s p e@(fvs, AnnVar v)
291 | isFollowableRep v_rep
292 = -- Ptr-ish thing; push it in the normal way
293 schemeT d s p (fvs, AnnVar v)
296 = -- returning an unboxed value. Heave it on the stack, SLIDE, and RETURN.
297 pushAtom True d p (AnnVar v) `thenBc` \ (push, szw) ->
298 returnBc (push -- value onto stack
299 `appOL` mkSLIDE szw (d-s) -- clear to sequel
300 `snocOL` RETURN v_rep) -- go
302 v_rep = typePrimRep (idType v)
304 schemeE d s p (fvs, AnnLit literal)
305 = pushAtom True d p (AnnLit literal) `thenBc` \ (push, szw) ->
306 let l_rep = literalPrimRep literal
307 in returnBc (push -- value onto stack
308 `appOL` mkSLIDE szw (d-s) -- clear to sequel
309 `snocOL` RETURN l_rep) -- go
313 Deal specially with the cases
314 let x = fn atom1 .. atomn in B
316 let x = Con atom1 .. atomn in B
317 (Con must be saturated)
319 In these cases, generate code to allocate in-line.
321 This is optimisation of the general case for let, which follows
322 this one; this case can safely be omitted. The reduction in
323 interpreter execution time seems to be around 5% for some programs,
324 with a similar drop in allocations.
326 This optimisation should be done more cleanly. As-is, it is
327 inapplicable to RHSs in letrecs, and needlessly duplicates code in
328 schemeR and schemeT. Some refactoring of the machinery would cure
331 schemeE d s p ee@(fvs, AnnLet (AnnNonRec x rhs) b)
333 = let d_init = if is_con then d else d'
335 mkPushes d_init args_r_to_l_reordered `thenBc` \ (d_final, push_code) ->
336 schemeE d' s p' b `thenBc` \ body_code ->
337 let size = d_final - d_init
338 alloc = if is_con then nilOL else unitOL (ALLOC size)
339 pack = unitOL (if is_con then PACK the_dcon size else MKAP size size)
341 returnBc (alloc `appOL` push_code `appOL` pack
344 -- Decide whether we can do this or not
345 (ok_to_go, is_con, the_dcon, the_fn)
347 Nothing -> (False, bomb 1, bomb 2, bomb 3)
348 Just (Left fn) -> (True, False, bomb 5, fn)
350 | dataConRepArity dcon <= length args_r_to_l
351 -> (True, True, dcon, bomb 6)
353 -> (False, bomb 7, bomb 8, bomb 9)
354 bomb n = panic ("schemeE.is_con(hacky hack hack) " ++ show n)
356 -- Extract the args (R -> L) and fn
357 args_r_to_l_reordered
361 = filter (not.isPtr.snd) args_r_to_l ++ filter (isPtr.snd) args_r_to_l
362 where isPtr = isFollowableRep . atomRep
364 args_r_to_l = filter (not.isTypeAtom.snd) args_r_to_l_raw
365 isTypeAtom (AnnType _) = True
368 (args_r_to_l_raw, maybe_fn) = chomp rhs
372 | isFCallId v || isPrimOpId v
375 -> case isDataConId_maybe v of
376 Just dcon -> ([], Just (Right dcon))
377 Nothing -> ([], Just (Left v))
378 AnnApp f a -> case chomp f of (az, f) -> (a:az, f)
379 AnnNote n e -> chomp e
380 other -> ([], Nothing)
382 -- This is the env in which to translate the body
386 -- Shove the args on the stack, including the fn in the non-dcon case
387 tag_when_push = not is_con
389 mkPushes :: Int{-curr depth-} -> [AnnExpr Id VarSet]
390 -> BcM (Int{-final depth-}, BCInstrList)
393 = returnBc (dd, nilOL)
395 = pushAtom False dd p' (AnnVar the_fn) `thenBc` \ (fn_push_code, fn_szw) ->
396 returnBc (dd+fn_szw, fn_push_code)
397 mkPushes dd (atom:atoms)
398 = pushAtom tag_when_push dd p' (snd atom)
399 `thenBc` \ (push1_code, push1_szw) ->
400 mkPushes (dd+push1_szw) atoms `thenBc` \ (dd_final, push_rest) ->
401 returnBc (dd_final, push1_code `appOL` push_rest)
404 -- General case for let. Generates correct, if inefficient, code in
406 schemeE d s p (fvs, AnnLet binds b)
407 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
408 AnnRec xs_n_rhss -> unzip xs_n_rhss
411 is_local id = not (isTyVar id) && elemFM id p'
412 fvss = map (filter is_local . varSetElems . fst) rhss
414 -- Sizes of tagged free vars, + 1 for the fn
415 sizes = map (\rhs_fvs -> 1 + sum (map taggedIdSizeW rhs_fvs)) fvss
417 -- This p', d' defn is safe because all the items being pushed
418 -- are ptrs, so all have size 1. d' and p' reflect the stack
419 -- after the closures have been allocated in the heap (but not
420 -- filled in), and pointers to them parked on the stack.
421 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n 1)))
424 infos = zipE4 fvss sizes xs [n, n-1 .. 1]
425 zipE = zipEqual "schemeE"
426 zipE4 = zipWith4Equal "schemeE" (\a b c d -> (a,b,c,d))
428 -- ToDo: don't build thunks for things with no free variables
429 buildThunk dd ([], size, id, off)
430 = returnBc (PUSH_G (Left (getName id))
431 `consOL` unitOL (MKAP (off+size-1) size))
432 buildThunk dd ((fv:fvs), size, id, off)
433 = pushAtom True dd p' (AnnVar fv)
434 `thenBc` \ (push_code, pushed_szw) ->
435 buildThunk (dd+pushed_szw) (fvs, size, id, off)
436 `thenBc` \ more_push_code ->
437 returnBc (push_code `appOL` more_push_code)
439 genThunkCode = mapBc (buildThunk d') infos `thenBc` \ tcodes ->
440 returnBc (concatOL tcodes)
442 allocCode = toOL (map ALLOC sizes)
444 schemeRs [] _ _ = returnBc ()
445 schemeRs (fvs:fvss) (x:xs) (rhs:rhss) =
446 schemeR False fvs (x,rhs) `thenBc_` schemeRs fvss xs rhss
448 schemeE d' s p' b `thenBc` \ bodyCode ->
449 schemeRs fvss xs rhss `thenBc_`
450 genThunkCode `thenBc` \ thunkCode ->
451 returnBc (allocCode `appOL` thunkCode `appOL` bodyCode)
457 schemeE d s p (fvs_case, AnnCase (fvs_scrut, scrut) bndr
458 [(DEFAULT, [], (fvs_rhs, rhs))])
460 | let isFunType var_type
461 = case splitTyConApp_maybe var_type of
462 Just (tycon,_) | isFunTyCon tycon -> True
464 ty_bndr = repType (idType bndr)
465 in isFunType ty_bndr || isTyVarTy ty_bndr
468 -- case scrut::suspect of bndr { DEFAULT -> rhs }
470 -- let bndr = scrut in rhs
471 -- when suspect is polymorphic or arrowtyped
472 -- So the required strictness properties are not observed.
473 -- At some point, must fix this properly.
477 (AnnNonRec bndr (fvs_scrut, scrut)) (fvs_rhs, rhs)
480 in trace ("WARNING: ignoring polymorphic case in interpreted mode.\n" ++
481 " Possibly due to strict polymorphic/functional constructor args.\n" ++
482 " Your program may leak space unexpectedly.\n")
483 (schemeE d s p new_expr)
487 {- Convert case .... of (# VoidRep'd-thing, a #) -> ...
489 case .... of a -> ...
490 Use a as the name of the binder too.
492 Also case .... of (# a #) -> ...
494 case .... of a -> ...
496 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
497 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
498 = --trace "automagic mashing of case alts (# VoidRep, a #)" (
499 schemeE d s p (fvs, AnnCase scrut bind2 [(DEFAULT, [bind2], rhs)])
502 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
503 | isUnboxedTupleCon dc
504 = --trace "automagic mashing of case alts (# a #)" (
505 schemeE d s p (fvs, AnnCase scrut bind1 [(DEFAULT, [bind1], rhs)])
508 schemeE d s p (fvs, AnnCase scrut bndr alts)
510 -- Top of stack is the return itbl, as usual.
511 -- underneath it is the pointer to the alt_code BCO.
512 -- When an alt is entered, it assumes the returned value is
513 -- on top of the itbl.
516 -- Env and depth in which to compile the alts, not including
517 -- any vars bound by the alts themselves
518 d' = d + ret_frame_sizeW + taggedIdSizeW bndr
519 p' = addToFM p bndr (d' - 1)
521 scrut_primrep = typePrimRep (idType bndr)
523 | scrut_primrep == PtrRep
525 | scrut_primrep `elem`
526 [CharRep, AddrRep, WordRep, IntRep, FloatRep, DoubleRep,
527 VoidRep, Int8Rep, Int16Rep, Int32Rep, Int64Rep,
528 Word8Rep, Word16Rep, Word32Rep, Word64Rep]
531 = pprPanic "ByteCodeGen.schemeE" (ppr scrut_primrep)
533 -- given an alt, return a discr and code for it.
534 codeAlt alt@(discr, binds_f, rhs)
536 = let (unpack_code, d_after_unpack, p_after_unpack)
537 = mkUnpackCode (filter (not.isTyVar) binds_f) d' p'
538 in schemeE d_after_unpack s p_after_unpack rhs
539 `thenBc` \ rhs_code ->
540 returnBc (my_discr alt, unpack_code `appOL` rhs_code)
542 = ASSERT(null binds_f)
543 schemeE d' s p' rhs `thenBc` \ rhs_code ->
544 returnBc (my_discr alt, rhs_code)
546 my_discr (DEFAULT, binds, rhs) = NoDiscr
547 my_discr (DataAlt dc, binds, rhs)
548 | isUnboxedTupleCon dc
549 = unboxedTupleException
551 = DiscrP (dataConTag dc - fIRST_TAG)
552 my_discr (LitAlt l, binds, rhs)
553 = case l of MachInt i -> DiscrI (fromInteger i)
554 MachFloat r -> DiscrF (fromRational r)
555 MachDouble r -> DiscrD (fromRational r)
556 MachChar i -> DiscrI i
557 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
560 | not isAlgCase = Nothing
562 = case [dc | (DataAlt dc, _, _) <- alts] of
564 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
567 mapBc codeAlt alts `thenBc` \ alt_stuff ->
568 mkMultiBranch maybe_ncons alt_stuff `thenBc` \ alt_final ->
570 alt_final_ac = ARGCHECK (taggedIdSizeW bndr) `consOL` alt_final
571 alt_bco_name = getName bndr
572 alt_bco = mkProtoBCO alt_bco_name alt_final_ac (Left alts)
574 schemeE (d + ret_frame_sizeW)
575 (d + ret_frame_sizeW) p scrut `thenBc` \ scrut_code ->
577 emitBc alt_bco `thenBc_`
578 returnBc (PUSH_AS alt_bco_name scrut_primrep `consOL` scrut_code)
581 schemeE d s p (fvs, AnnNote note body)
585 = pprPanic "ByteCodeGen.schemeE: unhandled case"
586 (pprCoreExpr (deAnnotate other))
589 -- Compile code to do a tail call. Specifically, push the fn,
590 -- slide the on-stack app back down to the sequel depth,
591 -- and enter. Four cases:
594 -- An application "PrelGHC.tagToEnum# <type> unboxed-int".
595 -- The int will be on the stack. Generate a code sequence
596 -- to convert it to the relevant constructor, SLIDE and ENTER.
598 -- 1. A nullary constructor. Push its closure on the stack
599 -- and SLIDE and RETURN.
601 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
602 -- it simply as b -- since the representations are identical
603 -- (the VoidRep takes up zero stack space). Also, spot
604 -- (# b #) and treat it as b.
606 -- 3. The fn denotes a ccall. Defer to generateCCall.
608 -- 4. Application of a non-nullary constructor, by defn saturated.
609 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
610 -- then the ptrs, and then do PACK and RETURN.
612 -- 5. Otherwise, it must be a function call. Push the args
613 -- right to left, SLIDE and ENTER.
615 schemeT :: Int -- Stack depth
616 -> Sequel -- Sequel depth
617 -> BCEnv -- stack env
623 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
624 -- = panic "schemeT ?!?!"
626 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate app)) ++ "\n") False
630 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
631 = pushAtom True d p arg `thenBc` \ (push, arg_words) ->
632 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
633 returnBc (push `appOL` tagToId_sequence
634 `appOL` mkSLIDE 1 (d+arg_words-s)
638 | is_con_call && null args_r_to_l
640 (PUSH_G (Left (getName con)) `consOL` mkSLIDE 1 (d-s))
645 | [arg1,arg2] <- args_r_to_l,
646 let isVoidRepAtom (_, AnnVar v) = VoidRep == typePrimRep (idType v)
647 isVoidRepAtom (_, AnnNote n e) = isVoidRepAtom e
648 in isVoidRepAtom arg2
649 = --trace (if isSingleton args_r_to_l
650 -- then "schemeT: unboxed singleton"
651 -- else "schemeT: unboxed pair with Void first component") (
656 | Just (CCall ccall_spec) <- isFCallId_maybe fn
657 = generateCCall d s p ccall_spec fn args_r_to_l
661 = if is_con_call && isUnboxedTupleCon con
662 then unboxedTupleException
663 else do_pushery d (map snd args_final_r_to_l)
666 -- Detect and extract relevant info for the tagToEnum kludge.
667 maybe_is_tagToEnum_call
668 = let extract_constr_Names ty
669 = case splitTyConApp_maybe (repType ty) of
670 (Just (tyc, [])) | isDataTyCon tyc
671 -> map getName (tyConDataCons tyc)
672 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
675 (_, AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
676 -> case isPrimOpId_maybe v of
677 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
681 -- Extract the args (R->L) and fn
682 (args_r_to_l, fn) = chomp app
687 | isTypeAtom (snd a) -> chomp f
688 | otherwise -> case chomp f of (az, f) -> (a:az, f)
689 AnnNote n e -> chomp e
690 other -> pprPanic "schemeT"
691 (ppr (deAnnotate (panic "schemeT.chomp", other)))
693 n_args = length args_r_to_l
695 isTypeAtom (AnnType _) = True
698 -- decide if this is a constructor application, because we need
699 -- to rearrange the arguments on the stack if so. For building
700 -- a constructor, we put pointers before non-pointers and omit
703 -- Also if the constructor is not saturated, we just arrange to
704 -- call the curried worker instead.
706 maybe_dcon = case isDataConId_maybe fn of
707 Just con | dataConRepArity con == n_args -> Just con
709 is_con_call = isJust maybe_dcon
710 (Just con) = maybe_dcon
716 = filter (not.isPtr.snd) args_r_to_l ++ filter (isPtr.snd) args_r_to_l
717 where isPtr = isFollowableRep . atomRep
719 -- make code to push the args and then do the SLIDE-ENTER thing
720 tag_when_push = not is_con_call
721 narg_words = sum (map (get_arg_szw . atomRep . snd) args_r_to_l)
722 get_arg_szw = if tag_when_push then taggedSizeW else untaggedSizeW
724 do_pushery d (arg:args)
725 = pushAtom tag_when_push d p arg `thenBc` \ (push, arg_words) ->
726 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
727 returnBc (push `appOL` more_push_code)
729 | Just (CCall ccall_spec) <- isFCallId_maybe fn
730 = panic "schemeT.do_pushery: unexpected ccall"
733 Just con -> returnBc (
734 (PACK con narg_words `consOL`
735 mkSLIDE 1 (d - narg_words - s)) `snocOL`
739 -> pushAtom True d p (AnnVar fn)
740 `thenBc` \ (push, arg_words) ->
741 returnBc (push `appOL` mkSLIDE (narg_words+arg_words)
746 {- Deal with a CCall. Taggedly push the args onto the stack R->L,
747 deferencing ForeignObj#s and (ToDo: adjusting addrs to point to
748 payloads in Ptr/Byte arrays). Then, generate the marshalling
749 (machine) code for the ccall, and create bytecodes to call that and
750 then return in the right way.
752 generateCCall :: Int -> Sequel -- stack and sequel depths
754 -> CCallSpec -- where to call
755 -> Id -- of target, for type info
756 -> [AnnExpr Id VarSet] -- args (atoms)
759 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
762 addr_usizeW = untaggedSizeW AddrRep
763 addr_tsizeW = taggedSizeW AddrRep
765 -- Get the args on the stack, with tags and suitably
766 -- dereferenced for the CCall. For each arg, return the
767 -- depth to the first word of the bits for that arg, and the
768 -- PrimRep of what was actually pushed.
770 pargs d [] = returnBc []
772 = let rep_arg = atomRep a
774 -- Don't push the FO; instead push the Addr# it
777 -> pushAtom False{-irrelevant-} d p a
778 `thenBc` \ (push_fo, _) ->
779 let foro_szW = taggedSizeW ForeignObjRep
780 d_now = d + addr_tsizeW
781 code = push_fo `appOL` toOL [
782 UPK_TAG addr_usizeW 0 0,
783 SLIDE addr_tsizeW foro_szW
785 in pargs d_now az `thenBc` \ rest ->
786 returnBc ((code, AddrRep) : rest)
789 -> pargs (d + addr_tsizeW) az `thenBc` \ rest ->
790 parg_ArrayishRep arrPtrsHdrSize d p a
792 returnBc ((code,AddrRep):rest)
795 -> pargs (d + addr_tsizeW) az `thenBc` \ rest ->
796 parg_ArrayishRep arrWordsHdrSize d p a
798 returnBc ((code,AddrRep):rest)
800 -- Default case: push taggedly, but otherwise intact.
802 -> pushAtom True d p a `thenBc` \ (code_a, sz_a) ->
803 pargs (d+sz_a) az `thenBc` \ rest ->
804 returnBc ((code_a, rep_arg) : rest)
806 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
807 -- the stack but then advance it over the headers, so as to
808 -- point to the payload.
809 parg_ArrayishRep hdrSizeW d p a
810 = pushAtom False{-irrel-} d p a `thenBc` \ (push_fo, _) ->
811 -- The ptr points at the header. Advance it over the
812 -- header and then pretend this is an Addr# (push a tag).
813 returnBc (push_fo `snocOL`
814 SWIZZLE 0 (hdrSizeW * untaggedSizeW PtrRep
817 PUSH_TAG addr_usizeW)
820 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
822 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
824 push_args = concatOL pushs_arg
825 d_after_args = d0 + sum (map taggedSizeW a_reps_pushed_r_to_l)
827 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
828 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
830 = reverse (tail a_reps_pushed_r_to_l)
832 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
833 -- push_args is the code to do that.
834 -- d_after_args is the stack depth once the args are on.
836 -- Get the result rep.
837 (returns_void, r_rep)
838 = case maybe_getCCallReturnRep (idType fn) of
839 Nothing -> (True, VoidRep)
840 Just rr -> (False, rr)
842 Because the Haskell stack grows down, the a_reps refer to
843 lowest to highest addresses in that order. The args for the call
844 are on the stack. Now push an unboxed, tagged Addr# indicating
845 the C function to call. Then push a dummy placeholder for the
846 result. Finally, emit a CCALL insn with an offset pointing to the
847 Addr# just pushed, and a literal field holding the mallocville
848 address of the piece of marshalling code we generate.
849 So, just prior to the CCALL insn, the stack looks like this
850 (growing down, as usual):
855 Addr# address_of_C_fn
856 <placeholder-for-result#> (must be an unboxed type)
858 The interpreter then calls the marshall code mentioned
859 in the CCALL insn, passing it (& <placeholder-for-result#>),
860 that is, the addr of the topmost word in the stack.
861 When this returns, the placeholder will have been
862 filled in. The placeholder is slid down to the sequel
863 depth, and we RETURN.
865 This arrangement makes it simple to do f-i-dynamic since the Addr#
866 value is the first arg anyway. It also has the virtue that the
867 stack is GC-understandable at all times.
869 The marshalling code is generated specifically for this
870 call site, and so knows exactly the (Haskell) stack
871 offsets of the args, fn address and placeholder. It
872 copies the args to the C stack, calls the stacked addr,
873 and parks the result back in the placeholder. The interpreter
874 calls it as a normal C call, assuming it has a signature
875 void marshall_code ( StgWord* ptr_to_top_of_stack )
877 -- resolve static address
881 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
883 -> let sym_to_find = _UNPK_ target in
884 ioToBc (lookupSymbol sym_to_find) `thenBc` \res ->
886 Just aa -> returnBc (True, aa)
887 Nothing -> ioToBc (linkFail "ByteCodeGen.generateCCall"
890 -> pprPanic "ByteCodeGen.generateCCall: casm" (ppr ccall_spec)
892 get_target_info `thenBc` \ (is_static, static_target_addr) ->
895 -- Get the arg reps, zapping the leading Addr# in the dynamic case
896 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
897 | is_static = a_reps_pushed_RAW
898 | otherwise = if null a_reps_pushed_RAW
899 then panic "ByteCodeGen.generateCCall: dyn with no args"
900 else tail a_reps_pushed_RAW
903 (push_Addr, d_after_Addr)
905 = (toOL [PUSH_UBX (Right static_target_addr) addr_usizeW,
906 PUSH_TAG addr_usizeW],
907 d_after_args + addr_tsizeW)
908 | otherwise -- is already on the stack
909 = (nilOL, d_after_args)
911 -- Push the return placeholder. For a call returning nothing,
912 -- this is a VoidRep (tag).
913 r_usizeW = untaggedSizeW r_rep
914 r_tsizeW = taggedSizeW r_rep
915 d_after_r = d_after_Addr + r_tsizeW
916 r_lit = mkDummyLiteral r_rep
917 push_r = (if returns_void
919 else unitOL (PUSH_UBX (Left r_lit) r_usizeW))
921 unitOL (PUSH_TAG r_usizeW)
923 -- generate the marshalling code we're going to call
926 arg1_offW = r_tsizeW + addr_tsizeW
927 args_offW = map (arg1_offW +)
928 (init (scanl (+) 0 (map taggedSizeW a_reps)))
930 ioToBc (mkMarshalCode cconv
931 (r_offW, r_rep) addr_offW
932 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
933 recordMallocBc addr_of_marshaller `thenBc_`
936 do_call = unitOL (CCALL (castPtr addr_of_marshaller))
938 wrapup = mkSLIDE r_tsizeW (d_after_r - r_tsizeW - s)
939 `snocOL` RETURN r_rep
941 --trace (show (arg1_offW, args_offW , (map taggedSizeW a_reps) )) (
944 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
949 -- Make a dummy literal, to be used as a placeholder for FFI return
950 -- values on the stack.
951 mkDummyLiteral :: PrimRep -> Literal
954 CharRep -> MachChar 0
956 WordRep -> MachWord 0
957 DoubleRep -> MachDouble 0
958 FloatRep -> MachFloat 0
959 AddrRep | taggedSizeW AddrRep == taggedSizeW WordRep -> MachWord 0
960 _ -> moan64 "mkDummyLiteral" (ppr pr)
964 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
965 -- -> (# PrelGHC.State# PrelGHC.RealWorld, PrelGHC.Int# #)
968 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
970 -- Alternatively, for call-targets returning nothing, convert
972 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
973 -- -> (# PrelGHC.State# PrelGHC.RealWorld #)
977 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
978 maybe_getCCallReturnRep fn_ty
979 = let (a_tys, r_ty) = splitRepFunTys fn_ty
981 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
983 = case splitTyConApp_maybe (repType r_ty) of
984 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
986 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
987 || r_reps == [VoidRep] )
988 && isUnboxedTupleTyCon r_tycon
989 && case maybe_r_rep_to_go of
991 Just r_rep -> r_rep /= PtrRep
992 -- if it was, it would be impossible
993 -- to create a valid return value
994 -- placeholder on the stack
995 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
998 --trace (showSDoc (ppr (a_reps, r_reps))) (
999 if ok then maybe_r_rep_to_go else blargh
1002 atomRep (AnnVar v) = typePrimRep (idType v)
1003 atomRep (AnnLit l) = literalPrimRep l
1004 atomRep (AnnNote n b) = atomRep (snd b)
1005 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1006 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1007 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1010 -- Compile code which expects an unboxed Int on the top of stack,
1011 -- (call it i), and pushes the i'th closure in the supplied list
1012 -- as a consequence.
1013 implement_tagToId :: [Name] -> BcM BCInstrList
1014 implement_tagToId names
1015 = ASSERT(not (null names))
1016 getLabelsBc (length names) `thenBc` \ labels ->
1017 getLabelBc `thenBc` \ label_fail ->
1018 getLabelBc `thenBc` \ label_exit ->
1019 zip4 labels (tail labels ++ [label_fail])
1020 [0 ..] names `bind` \ infos ->
1021 map (mkStep label_exit) infos `bind` \ steps ->
1022 returnBc (concatOL steps
1024 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1026 mkStep l_exit (my_label, next_label, n, name_for_n)
1027 = toOL [LABEL my_label,
1028 TESTEQ_I n next_label,
1029 PUSH_G (Left name_for_n),
1033 -- Make code to unpack the top-of-stack constructor onto the stack,
1034 -- adding tags for the unboxed bits. Takes the PrimReps of the
1035 -- constructor's arguments. off_h and off_s are travelling offsets
1036 -- along the constructor and the stack.
1038 -- Supposing a constructor in the heap has layout
1040 -- Itbl p_1 ... p_i np_1 ... np_j
1042 -- then we add to the stack, shown growing down, the following:
1054 -- so that in the common case (ptrs only) a single UNPACK instr can
1055 -- copy all the payload of the constr onto the stack with no further ado.
1057 mkUnpackCode :: [Id] -- constr args
1058 -> Int -- depth before unpack
1059 -> BCEnv -- env before unpack
1060 -> (BCInstrList, Int, BCEnv)
1061 mkUnpackCode vars d p
1062 = --trace ("mkUnpackCode: " ++ showSDocDebug (ppr vars)
1063 -- ++ " --> " ++ show d' ++ "\n" ++ showSDocDebug (ppBCEnv p')
1065 (code_p `appOL` code_np, d', p')
1069 vreps = [(var, typePrimRep (idType var)) | var <- vars]
1071 -- ptrs and nonptrs, forward
1072 vreps_p = filter (isFollowableRep.snd) vreps
1073 vreps_np = filter (not.isFollowableRep.snd) vreps
1075 -- the order in which we will augment the environment
1076 vreps_env = reverse vreps_p ++ reverse vreps_np
1078 -- new env and depth
1079 vreps_env_tszsw = map (taggedSizeW.snd) vreps_env
1080 p' = addListToFM p (zip (map fst vreps_env)
1081 (mkStackOffsets d vreps_env_tszsw))
1082 d' = d + sum vreps_env_tszsw
1084 -- code to unpack the ptrs
1085 ptrs_szw = sum (map (untaggedSizeW.snd) vreps_p)
1086 code_p | null vreps_p = nilOL
1087 | otherwise = unitOL (UNPACK ptrs_szw)
1089 -- code to unpack the nonptrs
1090 vreps_env_uszw = sum (map (untaggedSizeW.snd) vreps_env)
1091 code_np = do_nptrs vreps_env_uszw ptrs_szw (reverse (map snd vreps_np))
1092 do_nptrs off_h off_s [] = nilOL
1093 do_nptrs off_h off_s (npr:nprs)
1094 | npr `elem` [IntRep, WordRep, FloatRep, DoubleRep,
1095 CharRep, AddrRep, StablePtrRep]
1098 = moan64 "ByteCodeGen.mkUnpackCode" (ppr npr)
1100 approved = UPK_TAG usizeW (off_h-usizeW) off_s `consOL` theRest
1101 theRest = do_nptrs (off_h-usizeW) (off_s + tsizeW) nprs
1102 usizeW = untaggedSizeW npr
1103 tsizeW = taggedSizeW npr
1106 -- Push an atom onto the stack, returning suitable code & number of
1107 -- stack words used. Pushes it either tagged or untagged, since
1108 -- pushAtom is used to set up the stack prior to copying into the
1109 -- heap for both APs (requiring tags) and constructors (which don't).
1111 -- NB this means NO GC between pushing atoms for a constructor and
1112 -- copying them into the heap. It probably also means that
1113 -- tail calls MUST be of the form atom{atom ... atom} since if the
1114 -- expression head was allowed to be arbitrary, there could be GC
1115 -- in between pushing the arg atoms and completing the head.
1116 -- (not sure; perhaps the allocate/doYouWantToGC interface means this
1117 -- isn't a problem; but only if arbitrary graph construction for the
1118 -- head doesn't leave this BCO, since GC might happen at the start of
1119 -- each BCO (we consult doYouWantToGC there).
1121 -- Blargh. JRS 001206
1123 -- NB (further) that the env p must map each variable to the highest-
1124 -- numbered stack slot for it. For example, if the stack has depth 4
1125 -- and we tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1126 -- the tag in stack[5], the stack will have depth 6, and p must map v to
1127 -- 5 and not to 4. Stack locations are numbered from zero, so a depth
1128 -- 6 stack has valid words 0 .. 5.
1130 pushAtom :: Bool -> Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1131 pushAtom tagged d p (AnnVar v)
1133 | idPrimRep v == VoidRep
1134 = if tagged then returnBc (unitOL (PUSH_TAG 0), 1)
1135 else panic "ByteCodeGen.pushAtom(VoidRep,untaggedly)"
1138 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1140 | Just primop <- isPrimOpId_maybe v
1141 = returnBc (unitOL (PUSH_G (Right primop)), 1)
1145 str = "\npushAtom " ++ showSDocDebug (ppr v)
1146 ++ " :: " ++ showSDocDebug (pprType (idType v))
1147 ++ ", depth = " ++ show d
1148 ++ ", tagged = " ++ show tagged ++ ", env =\n" ++
1149 showSDocDebug (ppBCEnv p)
1150 ++ " --> words: " ++ show (snd result) ++ "\n" ++
1151 showSDoc (nest 4 (vcat (map ppr (fromOL (fst result)))))
1152 ++ "\nendPushAtom " ++ showSDocDebug (ppr v)
1156 = case lookupBCEnv_maybe p v of
1157 Just d_v -> (toOL (nOfThem nwords (PUSH_L (d-d_v+sz_t-2))), nwords)
1158 Nothing -> ASSERT(sz_t == 1) (unitOL (PUSH_G (Left nm)), nwords)
1160 nm = case isDataConId_maybe v of
1162 Nothing -> getName v
1164 sz_t = taggedIdSizeW v
1165 sz_u = untaggedIdSizeW v
1166 nwords = if tagged then sz_t else sz_u
1170 pushAtom True d p (AnnLit lit)
1171 = pushAtom False d p (AnnLit lit) `thenBc` \ (ubx_code, ubx_size) ->
1172 returnBc (ubx_code `snocOL` PUSH_TAG ubx_size, 1 + ubx_size)
1174 pushAtom False d p (AnnLit lit)
1176 MachLabel fs -> code CodePtrRep
1177 MachWord w -> code WordRep
1178 MachInt i -> code IntRep
1179 MachFloat r -> code FloatRep
1180 MachDouble r -> code DoubleRep
1181 MachChar c -> code CharRep
1182 MachStr s -> pushStr s
1185 = let size_host_words = untaggedSizeW rep
1186 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1190 = let getMallocvilleAddr
1192 CharStr s i -> returnBc (Ptr s)
1194 FastString _ l ba ->
1195 -- sigh, a string in the heap is no good to us.
1196 -- We need a static C pointer, since the type of
1197 -- a string literal is Addr#. So, copy the string
1198 -- into C land and introduce a memory leak
1199 -- at the same time.
1201 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1202 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1203 recordMallocBc ptr `thenBc_`
1205 do memcpy ptr ba (fromIntegral n)
1206 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1209 other -> panic "ByteCodeGen.pushAtom.pushStr"
1211 getMallocvilleAddr `thenBc` \ addr ->
1212 -- Get the addr on the stack, untaggedly
1213 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1219 pushAtom tagged d p (AnnApp f (_, AnnType _))
1220 = pushAtom tagged d p (snd f)
1222 pushAtom tagged d p (AnnNote note e)
1223 = pushAtom tagged d p (snd e)
1225 pushAtom tagged d p (AnnLam x e)
1227 = pushAtom tagged d p (snd e)
1229 pushAtom tagged d p other
1230 = pprPanic "ByteCodeGen.pushAtom"
1231 (pprCoreExpr (deAnnotate (undefined, other)))
1233 foreign import "memcpy" memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1236 -- Given a bunch of alts code and their discrs, do the donkey work
1237 -- of making a multiway branch using a switch tree.
1238 -- What a load of hassle!
1239 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1240 -- a hint; generates better code
1241 -- Nothing is always safe
1242 -> [(Discr, BCInstrList)]
1244 mkMultiBranch maybe_ncons raw_ways
1245 = let d_way = filter (isNoDiscr.fst) raw_ways
1246 notd_ways = naturalMergeSortLe
1247 (\w1 w2 -> leAlt (fst w1) (fst w2))
1248 (filter (not.isNoDiscr.fst) raw_ways)
1250 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1251 mkTree [] range_lo range_hi = returnBc the_default
1253 mkTree [val] range_lo range_hi
1254 | range_lo `eqAlt` range_hi
1255 = returnBc (snd val)
1257 = getLabelBc `thenBc` \ label_neq ->
1258 returnBc (mkTestEQ (fst val) label_neq
1260 `appOL` unitOL (LABEL label_neq)
1261 `appOL` the_default))
1263 mkTree vals range_lo range_hi
1264 = let n = length vals `div` 2
1265 vals_lo = take n vals
1266 vals_hi = drop n vals
1267 v_mid = fst (head vals_hi)
1269 getLabelBc `thenBc` \ label_geq ->
1270 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1271 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1272 returnBc (mkTestLT v_mid label_geq
1274 `appOL` unitOL (LABEL label_geq)
1278 = case d_way of [] -> unitOL CASEFAIL
1281 -- None of these will be needed if there are no non-default alts
1282 (mkTestLT, mkTestEQ, init_lo, init_hi)
1284 = panic "mkMultiBranch: awesome foursome"
1286 = case fst (head notd_ways) of {
1287 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1288 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1291 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1292 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1295 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1296 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1299 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1300 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1302 DiscrP algMaxBound )
1305 (algMinBound, algMaxBound)
1306 = case maybe_ncons of
1307 Just n -> (0, n - 1)
1308 Nothing -> (minBound, maxBound)
1310 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1311 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1312 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1313 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1314 NoDiscr `eqAlt` NoDiscr = True
1317 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1318 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1319 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1320 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1321 NoDiscr `leAlt` NoDiscr = True
1324 isNoDiscr NoDiscr = True
1327 dec (DiscrI i) = DiscrI (i-1)
1328 dec (DiscrP i) = DiscrP (i-1)
1329 dec other = other -- not really right, but if you
1330 -- do cases on floating values, you'll get what you deserve
1332 -- same snotty comment applies to the following
1334 minD, maxD :: Double
1340 mkTree notd_ways init_lo init_hi
1344 %************************************************************************
1346 \subsection{Supporting junk for the compilation schemes}
1348 %************************************************************************
1352 -- Describes case alts
1360 instance Outputable Discr where
1361 ppr (DiscrI i) = int i
1362 ppr (DiscrF f) = text (show f)
1363 ppr (DiscrD d) = text (show d)
1364 ppr (DiscrP i) = int i
1365 ppr NoDiscr = text "DEF"
1368 -- Find things in the BCEnv (the what's-on-the-stack-env)
1369 -- See comment preceding pushAtom for precise meaning of env contents
1370 --lookupBCEnv :: BCEnv -> Id -> Int
1371 --lookupBCEnv env nm
1372 -- = case lookupFM env nm of
1373 -- Nothing -> pprPanic "lookupBCEnv"
1374 -- (ppr nm $$ char ' ' $$ vcat (map ppr (fmToList env)))
1377 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1378 lookupBCEnv_maybe = lookupFM
1381 taggedIdSizeW, untaggedIdSizeW :: Id -> Int
1382 taggedIdSizeW = taggedSizeW . typePrimRep . idType
1383 untaggedIdSizeW = untaggedSizeW . typePrimRep . idType
1385 unboxedTupleException :: a
1386 unboxedTupleException
1389 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1390 "\tto foreign import/export decls in source. Workaround:\n" ++
1391 "\tcompile this module to a .o file, then restart session."))
1394 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1399 %************************************************************************
1401 \subsection{The bytecode generator's monad}
1403 %************************************************************************
1407 = BcM_State { bcos :: [ProtoBCO Name], -- accumulates completed BCOs
1408 nextlabel :: Int, -- for generating local labels
1409 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1410 -- Should be free()d when it is GCd
1411 type BcM r = BcM_State -> IO (BcM_State, r)
1413 ioToBc :: IO a -> BcM a
1414 ioToBc io st = do x <- io
1417 runBc :: BcM_State -> BcM r -> IO (BcM_State, r)
1418 runBc st0 m = do (st1, res) <- m st0
1421 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1422 thenBc expr cont st0
1423 = do (st1, q) <- expr st0
1424 (st2, r) <- cont q st1
1427 thenBc_ :: BcM a -> BcM b -> BcM b
1428 thenBc_ expr cont st0
1429 = do (st1, q) <- expr st0
1430 (st2, r) <- cont st1
1433 returnBc :: a -> BcM a
1434 returnBc result st = return (st, result)
1437 mapBc :: (a -> BcM b) -> [a] -> BcM [b]
1438 mapBc f [] = returnBc []
1440 = f x `thenBc` \ r ->
1441 mapBc f xs `thenBc` \ rs ->
1444 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM ()
1446 = return (st{bcos = bco (malloced st) : bcos st, malloced=[]}, ())
1450 | not (null (malloced st))
1451 = panic "ByteCodeGen.newbcoBc: missed prior emitBc?"
1455 recordMallocBc :: Ptr a -> BcM ()
1457 = return (st{malloced = castPtr a : malloced st}, ())
1459 getLabelBc :: BcM Int
1461 = return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1463 getLabelsBc :: Int -> BcM [Int]
1465 = let ctr = nextlabel st
1466 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])